Research map / Image generation / Text-to-image diffusion
Concept Removal Guidance: Evidence-Calibrated Negative Guidance for Safe Diffusion Sampling
Yoonseok Choi, Chaeyoung Oh, Hyunjun Choi, Seokin Seo, Kee-Eung Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adapt negative guidance with CRG instead of fixed-scale negative prompting | SD v1.4 · Nudity removal, P4D | ASR: 0.338 -> 0.026 (NP -> CRG, P4D 0.338 vs 0.026) | Improved |
| use CRG instead of DNG for nudity removal | SD v1.4 · COCO benign | FID: 55.97 -> 36.05 | Improved |
| use CRG instead of DNG on MMA-Diff | SD v1.4 · Nudity removal, MMA-Diff | ASR: 0.139 -> 0.164 | Hurt |
| apply CRG on SDXL and SD v3 backbones | CRG · Nudity removal, P4D (SDXL) | ASR: 0.748 -> 0.060 | Improved |
| use deterministic DPM-Solver++ sampler with CRG | CRG · Nudity removal, P4D (DPM-Solver++) | ASR: 0.960 -> 0.020 | Improved |
| apply CRG to violence removal | SD v1.4 · Violence removal, Ring-A-Bell 5-77 | ASR: 0.992 -> 0.296 | Improved |
| apply CRG to artist style removal | SD v1.4 · Artist style removal, Van Gogh target | Acc_e: 0.80 -> 0.10 | Improved |
| increase posterior variance factor k from 2 to 16 (look-ahead term) (ablation) | CRG · Nudity removal, MMA-Diff | ASR: 0.390 -> 0.164 | Improved |
| increase k beyond 16 (32, 64) (ablation) | CRG · Nudity removal, P4D | ASR: 0.026 -> 0.053 (k=32), 0.026 (k=64) | No clear effect |
| lower threshold tau from 30 to 0 (ablation) | CRG · Nudity removal (COCO benign / MMA-Diff) | ASR and FID: MMA-Diff ASR 0.164 -> 0.079; COCO FID 36.05 -> 52.05 | Mixed |
| increase guidance scale lambda0 from 1.4 to 2.2 (ablation) | CRG · Nudity removal, MMA-Diff | ASR and FID: ASR 0.417 -> 0.083; FID 32.47 -> 40.43 | Mixed |
| increase DNG temperature tau_temp from 0.5 to 2.0 (ablation) | DNG · Nudity removal, P4D | ASR and FID: ASR 0.517 -> 0.126; FID 53.55 -> 57.25 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SAFREE — SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Dynamic Negative Guidance of Diffusion Models (ICLR 2025 · similarity 0.90)
- Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models (ICLR 2025 · similarity 0.89)
- Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization (CVPR 2025 · similarity 0.89)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
